On (10/18/17 17:04), Tobin C. Harding wrote:
[..]
> > > +/* protects ptr_secret and have_key */
> > > +DEFINE_SPINLOCK(key_lock);
> > > +static siphash_key_t ptr_secret __read_mostly;
> > > +static atomic_t have_key = ATOMIC_INIT(0);
> > > +
> > > +static int initialize_ptr_secret(void)
> > > +{
> > > + spin_lock(&key_lock);
> > > + if (atomic_read(&have_key) == 1)
> > > + goto unlock;
> > > +
> > > + get_random_bytes(&ptr_secret, sizeof(ptr_secret));
> > > + atomic_set(&have_key, 1);
> > > +
> > > +unlock:
> > > + spin_unlock(&key_lock);
> > > + return 0;
> > > +}
> >
> > is this spinlock legal? what happens if we are getting interrupted by NMI?
>
> I think we can do without the spinlock. I think I was already told that when
> I tried to put it [some where else] in v1.
>
> It's fun failing in public ;)
another note is that
printk()->vscnprintf()->get_random_bytes()->warn_unseeded_randomness()
causes a printk() recursion, but we should be fine now, we are in printk_safe
mode by the time we vscnprintf().
but a bigger problem might the following thing:
vscnprintf()
pointer()
ptr_to_id()
initialize_ptr_secret()
get_random_bytes()
_get_random_bytes()
extract_crng()
_extract_crng()
spin_lock_irqsave(&crng->lock, flags); <<<<<
this, once again, can deadlock. can it? just like before:
> > printk()
> > vprintk_emit()
> > vscnprintf()
> > pointer()
> > ptr_to_id()
> > initialize_ptr_secret()
> > spin_lock(&key_lock)
> >
> > ----> NMI
> >
> > printk()
> > printk_safe_log_store()
> > vscnprintf()
> > pointer()
> > ptr_to_id()
> > initialize_ptr_secret()
> > spin_lock(&key_lock) <<<<
-ss
